Non-temporal neural networks for predicting degradation trends of key wind-turbine gearbox components

Deep-learning models, which are widely used to predict degradation trends, are effective for automatic extraction of hidden features from monitoring data and complete learning of the evolutionary patterns of component performance. However, wind turbine monitoring systems often yield discontinuous data owing to short-term sensor abnormalities or unstable data transmission links, leading to difficulties in training existing deep-learning models and low prediction accuracy for the degradation trends of key gearbox components. This study proposes a nontemporal neural network (NTNNet) with multitask training for predicting the degradation trends of key wind-turbine components using discontinuous data. Multiple support and query sets are constructed by intercepting short time-series samples, and multitask training is performed to overcome model training difficulties in the case of discontinuous samples. A gated recurrent unit is constructed using the meta-learning framework, and parameters are updated through multiple recursions to enhance the characterization ability for key-component performance. The accuracy and superiority of the proposed method are validated using two datasets and compared to existing methods. The proposed method achieves degradation trend prediction accuracies of 4.61, 4.73, 4.61, and 3.05 for the four components of wind turbine gearboxes, respectively, meetingthe practical application requirements for the health assessment of wind turbine gearboxes.